(Abridged) We revisit the question of the nature of ULXs through a detailed investigation of their spectral shape, using the highest quality X-ray data available in the XMM-Newton public archives. We confirm that simple spectral models commonly used for the analysis and interpretation of ULXs (power-law continuum and multi-colour disc blackbody models) are inadequate in the face of such high quality data. Instead we find two near ubiquitous features in the spectrum: a soft excess and a roll-over in the spectrum at energies above 3keV. We investigate a range of more physical models to describe these data. We find that disc plus Comptonised corona models fit the data well, but the derived corona is cool, and optically thick (tau ~ 5-30). We argue that these observed disc temperatures are not a good indicator of the black hole mass as the powerful, optically thick corona drains energy from the inner disc, and obscures it. We estimate the intrinsic (corona-less) disc temperature, and demonstrate that in most cases it lies in the regime of stellar mass black holes. These objects have spectra which range from those similar to the highest mass accretion rate states in Galactic binaries, to those which clearly have two peaks, one at energies below 1 keV (from the outer, unComptonised disc) and one above 3 keV (from the Comptonised, inner disc). However, a few ULXs have a significantly cooler corrected disc temperature; we suggest that these are the most extreme stellar mass black hole accretors, in which a massive wind completely envelopes the inner disc regions, creating a cool photosphere. We conclude that ULXs provide us with an observational template for the transition between Eddington and super-Eddington accretion flows, with the latter occupying a new ultraluminous accretion state.